Deflating the Sharpe after a grid search (the winner is a max over N)

[ipezygj]

vectorbt makes it trivial to test thousands of parameter/asset combinations at once — which is exactly why the top result of a grid needs one extra step before you trust it.

When you take the best Sharpe out of a grid, that number is a maximum over N trials, not a single-test Sharpe. The expected best of N random strategies is already well above zero, so the grid's winner can look great and still sit inside the noise band of the search. That's what the Deflated Sharpe Ratio (Bailey & López de Prado) corrects for — and the nice thing about a grid is you don't have to assume how noisy the search was, you can measure it from the spread of Sharpes the grid already produced.

A few lines on top of a vectorbt grid — MIT tool, pip install numguard, no heavy deps:

import numpy as np, vectorbt as vbt
from numguard.backtest import deflated_sharpe, expected_max_sharpe

# ... your usual grid, e.g. a fast/slow MA or RSI sweep, producing a multi-column Portfolio `pf` ...
sharpes = pf.sharpe_ratio().dropna()          # one (annualized) Sharpe per combination
n       = int(sharpes.size)                    # how many combos you actually searched
best    = float(sharpes.max())                 # the winner you'd report
T       = len(pf.wrapper.index)                # sample length in bars

# use the grid's OWN dispersion of Sharpes as the trial spread — measured, not assumed (per-bar units):
sr_std = float(sharpes.std()) / np.sqrt(252)
bar    = expected_max_sharpe(n_trials=n, sr_std=sr_std) * np.sqrt(252)
v      = deflated_sharpe(sr=best/np.sqrt(252), T=T, n_trials=n, sr_std=sr_std)

print(f"raw best (ann): {best:.2f}")
print(f"selection bar : {bar:.2f} ann   <- what {n} random tries at your spread reach by luck")
print(f"DSR           : {v.dsr:.3f}   survives={v.survives}")

If the DSR is near zero while the raw Sharpe looked strong, the grid found a lucky maximum, not an edge. It turns a "which combo won?" plot into "did the winner clear its own search bar?".

There's also a fuller battery on a returns series (CSCV Probability of Backtest Overfitting across the candidates matrix, a same-bar look-ahead scan, autocorrelation-inflated Sharpe) if you want it. Repo: https://github.com/ipezygj/numguard — feedback welcome, especially on the vectorbt wiring.